Neuro-Symbolic Transformer Architecture for Behavior Monitoring in Autonomous Vehicles

Self-driving cars are being increasingly deployed in real-world environments, but their black-box decision-making raises concerns about trust, transparency, and safety. There is a significant research gap in connecting high-performance perception models with interpretable reasoning frameworks for the real-time monitoring of autonomous vehicles (AVs). This paper presents a new hybrid neuro-symbolic monitoring framework that combines deep transformer-based perception with symbolic reasoning to enable real-time, interpretable behavior monitoring in AVs. Our method processes multi-modal sensory data, such as camera images, LiDAR, and telemetry, through a multi-head transformer encoder. The encoded features are projected into a predicate space via a neuro-symbolic bridge, where rule-based logic is used to identify behavioral violations and produce human-readable explanations. We test the system on both simulated (CARLA) and real-world (nuScenes) datasets, achieving 92.8% accuracy, 94.1% SHAP interpretability, and 94.3% explanation trace match rate. Through various deployment case studies, including red light violations, unsafe lane changes, and speeding in residential areas, we demonstrate the framework’s reliability, traceability, and adherence to policies. The system operates in real-time and integrates smoothly into AV stacks through ROS2 and Docker, providing a scalable route to transparent and trustworthy AV deployments.